Fix a race condition in scheduling code in `FileSystemValueCheckerInf…

…erringAncestors`.

`FileSystemValueCheckerInferringAncestors` works on the premise that we schedule all leaf entries from the diff and work our way up the path from there. The code which starts parallel processing detects leaves by looking at how many children to process each entry has. That is only correct if no processing is happening.

Fix the race condition happening due to the fact that we schedule the processing in a stream detecting the leaves -- this means that the detection algorithm works in parallel with processing. Add a materialization step in between to detect all leaves before any processing starts.

PiperOrigin-RevId: 390417452

src/main/java/com/google/devtools/build/lib/skyframe/FileSystemValueCheckerInferringAncestors.java

@@ -165,6 +165,10 @@ private ImmutableDiff processEntries(int nThreads)
     ImmutableList<Future<?>> futures =
         nodeStates.entrySet().stream()
             .filter(e -> e.getValue().childrenToProcess.get() == 0)
+            // Materialize all leaves before scheduling them -- otherwise, we could race with the
+            // processing code which decrements childrenToProcess.
+            .collect(toImmutableList())
+            .stream()
             .map(
                 e ->
                     executor.submit(